Training tagger/chunker with Maxent as classifier ends with fail. To do this, there must be installed old version of scipy (10.1).
What's more, training tagger with Maxent as classifier ends with Memmory Error (conll 2002 as corpus)
This is actually a NLTK issue: nltk/nltk#349
As noted there, it's been fixed for the next release.